Page 54 - Haematologica April 2020

M.G. Gorelashvili et al.
to the vasculature,14,15 and have a diameter of up to 50 μm in mice and 50-100 μm in humans.13 Mature megakary- ocytes produce platelets and release them into the blood circulation in order to maintain constant platelet counts. In addition, they actively regulate HSC proliferation in both positive and negative manners.16-19 Recent studies revealed that most HSC are in close proximity to sinusoidal blood vessels.1,4,20 Likewise, more than 70% of megakaryocytes were found to be in contact with the BM vasculature.15 Besides this indirect correlation, at least a subset of megakaryocytes was found to be in close proximity to HSC.16,18,19 Moreover, megakaryocytes were shown to influence HSC quiescence via different cytokines, such as CXCL4,16 transforming growth factor β1 (TGFβ1)17 and thrombopoietin.18,21 However, very recently it was shown that liver-derived, and not megakaryocyte-derived, throm- bopoietin is required for HSC maintenance in the BM.22
Nevertheless, megakaryocyte activity in the intact medullary space and its interplay with other BM cells has gained great attention in the last decade; numerous in vitro investigations, based on two-dimensional (2D) cryosec- tions, and in vivo (two-photon) imaging studies have been reported.23 Intrinsic limitations of these methods such as loss of volume information, cutting artifacts or small field of view impair scientific models15 and treatment of patients.24 Recently, whole bone optical clearing and imag- ing have been established to overcome these limitations. Despite the significant advances in imaging technology, tools for correct quantitative analysis of the geometry and localization of megakaryocytes, HSC and other BM com- ponents are still in their infancy. As image segmentation is a complex and error-prone method, exact definition of the image-processing pipeline is of great importance. The recently developed machine learning toolkits25,26 are pow- erful complements to the portfolio, and allow for more comprehensive data analysis, and access to previously masked information. Successful segmented objects derived from complex microscopy data can be used for in silico analysis of cell distributions within the BM architec- ture as recently demonstrated.15,16,20
Modular toolkits in particular have been proven to be powerful, not only for image analysis, but also for struc- ture reconstruction as well as simulations of growth and organization.27,28 Unfortunately, these tools are not yet uni- versally applicable. Here, we developed and compared dif- ferent image processing pipelines and simulation scenarios for precise identification of megakaryocytes in three- dimensional (3D) light sheet fluorescence microscopy (LSFM) image stacks of uncut murine bones. Megakaryocytes have been described to have an impact, based on biochemical processes, on cell migration in the BM. However, the impact of increases in their number and size is unclear. To date, the only available technique for investigating cell migration in the BM is intravital imaging, which is hampered by the limited time during which measurements can be made because of the need for anes- thesia of the animals and accumulating phototoxicity,29 limited penetration depth30 and a relatively small field of view of typically 300-500 μm2 that does not allow obser- vation of the whole bone simultaneously. Here, computa- tional simulations represent an important complementing and well-controllable tool for elucidating underlying cell mechanisms.31-33 Typically, simulation studies use artificial meshes as templates due to the lack of experimental data or to minimize the computational effort. Unfortunately,
such simplified artificial templates for megakaryocytes and the vasculature can bias simulations and lead to mis- interpretations as we show in this study. Here, we demon- strate that using the segmented cell and vessel objects of true 3D images can overcome those limitations, providing a simulation framework that has the prerequisites to reflect the physiological situation optimally.
Methods
More methodological details are present in the Online Supplementary Material.
Mice
All animal experiments were approved by the district govern- ment of Lower Frankonia (Bezirksregierung Unterfranken). We used 8- to 12-week old C57BL/6JRj (Janvier Labs) or Lyz2GFP mice.34
Thrombocytopenia was induced by intravenous injection of rat anti-GPIbα35 (CD42b; 2.0 μg/g body weight Emfret Analytics, Eibelstadt, Germany;).
Two-photon intravital imaging
Lyz2GFP mice34 were anesthetized by intraperitoneal injection of medetomidine 0.5mg/g, midazolam 5mg/g and fentanyl 0.05 mg/g body weight. A 1-cm midline incision was made to expose the frontoparietal skull, while carefully avoiding damage to the bone tissue. The mouse was placed on a custom-designed metal stand equipped with a stereotactic holder to immobilize the head. BM vasculature was visualized by injection of bovine serum albu- min-Alexa546 (8 μg/g body weight) and anti-CD105 Alexa546 (0.6 μg/g body weight). Neutrophils were visualized by the endoge- nously expressed green fluorescent protein. Stacks were acquired at a frame rate of 1/min on an upright two-photon fluorescence microscope (TCS SP8 MP, Leica Microsystems, Wetzlar, Germany) equipped with a 25x water objective with a numerical aperture of 1.0. A tunable broad-band Ti:Sa laser (Chameleon, Coherent, Dieburg, Germany) was used at 780 nm to capture green fluorescent protein and Alexa546 fluorescence. For each mouse, three time series of z-stacks were recorded (20 min each, 1 z-stack/min) with a voxel size of 0.87 x 0.87 x 1 μm3. Details on image analysis are provided in the Online Supplementary Material.
Light sheet fluorescence microscopy image processing and segmentation
Image stacks were processed, visualized, and analyzed using FIJI,36 Ilastik 1.2.25 and Imaris® 8.4 (Bitplane AG, Zurich, Switzerland). The four different analysis pipelines (I-IV), shared the same image-preprocessing steps performed in FIJI and Imaris. In the first pipeline (I), the membrane algorithm was directly applied on megakaryocytes. The second, simple one-pass pipeline (II) utilized the Imaris cell soma algorithm with one seeding step, whereas the extended two-pass pipeline (III) employed two sub- sequent seeding steps for large and small megakaryocytes. Our customized pipeline (IV) corrected for fake invaginations by creat- ing virtual cell somata before applying two-step seeding. Bone and BM were identified using the pixel classification algorithm of Ilastik software, with results transferred to Imaris 8.4 for segmen- tation and further analysis. Details are provided in the Online Supplementary Material.
Static and dynamic computational simulations
Simulations of megakaryocyte distribution (static) and cell migration (dynamic) in the BM were performed by custom-writ- ten algorithms (Online Supplementary Figures S4 and S6) in Matlab
896
haematologica | 2020; 105(4)